ComfyUI server setup

The plugins do not run AI models themselves. They send work to a ComfyUI server that loads the models, executes them, and returns the results. This page is the canonical guide to standing that server up.

Architecture recap 

[ OFX host ] ──HTTP──▶ [ ComfyUI server ]
       │                       │
       └──── shared filesystem ┘
            (input + output EXR files)

The host machine and the ComfyUI machine can be the same physical computer, or separate machines on a network. In a studio setting, separate is the more common pattern — one or more GPU machines serve many artists.

You need three things:

  1. A ComfyUI installation with the right custom nodes installed.
  2. The model weights for each plugin you intend to use.
  3. A shared filesystem that both the host and the ComfyUI server can read and write.

1. Install ComfyUI

Follow the official ComfyUI installation guide.

If you want the fastest path on Windows — Python, PyTorch, ComfyUI, and ComfyUI Manager in one shot — ComfyUI-Easy-Install is a one-click installer that sets the whole stack up in a self-contained folder. It’s a convenient starting point; the custom nodes and model weights in the next sections still need to be added on top.

GPU recommendations by plugin

The dominant resource is GPU VRAM. Approximate ranges (lower numbers achievable with quantized variants and CPU offload):

Plugin Typical VRAM Minimum VRAM Notes
depth_da3 6–10 GB 4 GB Smaller variants fit consumer GPUs comfortably.
normal_crafter 20 GB @ 1024×576 6 GB @ 512×256 SVD-based; resolution-sensitive.
depth_crafter 26 GB @ 1024×576 8 GB with CPU offload Diffusion video; the heaviest standard run.
segmentation_sam3 4 GB 4 GB 24 GB+ recommended for 4K work.
matte_mama High (SVD-XT base) TBD Single-pass diffusion despite the SVD base.
matte_ma2 Moderate TBD Recurrent network — lighter than diffusion.
upscale_seedvr2 4–24+ GB 4 GB (GGUF Q4_K_M) 7B FP16 wants H100; 3B and quantized variants run on consumer cards.

A single 24 GB GPU (RTX 3090 / 4090 / A5000-class) can handle every plugin in this suite at standard production resolutions. Heavy work at 4K wants 48 GB+.

2. Install the custom nodes

ComfyUI itself does not ship the model code for these plugins — that lives in community custom node repositories. Install via ComfyUI Manager or by cloning into ComfyUI/custom_nodes/.

Plugin Required custom node
depth_da3 PozzettiAndrea/ComfyUI-DepthAnythingV3
normal_crafter AIWarper/ComfyUI-NormalCrafterWrapper
depth_crafter akatz-ai/ComfyUI-DepthCrafter-Nodes
segmentation_sam3 PozzettiAndrea/ComfyUI-SAM3
matte_mama okdalto/ComfyUI-VideoMaMa (depends on SAM3)
matte_ma2 spiritform/comfy-matanyone2 (depends on SAM3)
upscale_seedvr2 numz/ComfyUI-SeedVR2_VideoUpscaler

Restart ComfyUI after installing custom nodes. Verify that each new node appears in the ComfyUI workflow editor before continuing.

3. Download model weights

Each plugin ships a fixed ComfyUI workflow whose model-loader node decides where the weights come from. There are two patterns, and the difference matters for setup:

  • Auto-download — the loader node fetches the selected checkpoint to a ComfyUI/models/ subfolder on first use. Nothing to do up front; just expect a slow first render while it downloads.
  • Manual — the loader (CheckpointLoaderSimple) reads a file you must place in ComfyUI/models/checkpoints/ yourself. The render fails immediately if the file is missing.
Plugin Loader node (in the workflow) How weights arrive Where
depth_da3 DownloadAndLoadDepthAnythingV3Model Auto-download models/depthanything3/
depth_crafter DownloadAndLoadDepthCrafterModel Auto-download models/ (DepthCrafter + SVD base)
normal_crafter (none — wrapper-internal) Auto-download wrapper-managed (SVD base)
segmentation_sam3 CheckpointLoaderSimple Manual + gated models/checkpoints/
matte_mama CheckpointLoaderSimple + VideoMaMaPipelineLoader Manual (SAM3, gated) + auto (MaMa) models/checkpoints/ + pipeline dirs
matte_ma2 CheckpointLoaderSimple Manual + gated (SAM3) + node-managed (MatAnyone2) models/checkpoints/
upscale_seedvr2 SeedVR2LoadDiTModel + SeedVR2LoadVAEModel Auto-download models/SEEDVR2/

SAM3 is gated. SAM3 (facebook/sam3) backs segmentation_sam3 and is a shared dependency of both matte plugins (matte_mama, matte_ma2). It is a gated Hugging Face model: run hf auth login and accept the terms on the model page, then place the checkpoint in ComfyUI/models/checkpoints/. It will not auto-download.

Each plugin’s documentation page lists the exact weights, sizes, and Hugging Face sources. Pre-downloading avoids a multi-minute stall on the first render.

License notice. Model weights are governed by their upstream licenses. Several models in this suite are restricted to non-commercial use. See each plugin page and the release spec for the full breakdown.

4. Set up the shared input/output folders

The plugin writes input frames as EXR files into an input directory. The ComfyUI workflow loads from there, runs the model, and writes the output back as EXR into an output directory. The plugin then reads the output back.

For this to work:

  • Both the host machine and the ComfyUI server must be able to read and write the same physical location.
  • Both sides need to know the path under their own mount point.

Single machine

Both sides see the same paths natively. Configure both to e.g.:

/Users/<you>/comfyui-share/in/
/Users/<you>/comfyui-share/out/

In each plugin’s parameters, set Client Mount Path and Server Mount Path to the same value.

Networked: Linux ComfyUI server, macOS / Linux clients

Mount the server’s storage on each client via NFS or SMB.

Side Path
Linux ComfyUI server /mnt/comfyui-share/in/
macOS client (Finder mount) /Volumes/comfyui-share/in/

In the plugin, set:

  • Client Mount Path: /Volumes/comfyui-share
  • Server Mount Path: /mnt/comfyui-share

The plugin substitutes the server path into the workflow JSON so ComfyUI receives the path it can actually read.

Networked: Windows ComfyUI server

Use UNC paths for the server side:

Side Path
Windows ComfyUI server \\COMFYBOX\share\in\
macOS / Linux client (mount) /Volumes/share/in/ (or wherever)

In the plugin:

  • Client Mount Path: /Volumes/share
  • Server Mount Path: \\COMFYBOX\share

The plugin handles the backslash escaping for the workflow JSON automatically.

Folder layout the plugin creates

Inside the input and output directories, the plugin creates a project / workflow / frame hierarchy automatically:

<input-mount>/
  <project-name>/
    <workflow-name>/
      <frame-or-sequence-id>.exr
<output-mount>/
  <project-name>/
    <workflow-name>/
      <frame-or-sequence-id>.exr

You only need to provision the top-level mount points. Everything underneath is managed by the plugin.

5. Start ComfyUI 

python main.py --listen 0.0.0.0 --port 8188

--listen 0.0.0.0 lets remote hosts connect. Restrict via firewall to your trusted network. Match the --port to the port in your ComfyUI Server URL (step 6); 8188 is the ComfyUI default.

You do not need --input-directory / --output-directory. Those flags only relocate ComfyUI’s default in/out folders. AIFX never uses the defaults: it writes absolute paths (built from the Server Mount Path) straight into the workflow’s LoadEXR / SaveEXR nodes, and SaveEXR creates its output folder if missing. So the shared-folder wiring is governed entirely by the mount paths in step 4 / step 6 — not by ComfyUI’s launch flags. Setting those flags has no effect on where AIFX reads and writes.

6. Configure the plugin

In your OFX host, apply a plugin to a clip and set:

  • ComfyUI Server URL: http://<server-ip>:8188
  • Client Mount Path: the input/output share path on the host machine.
  • Server Mount Path: the same share path on the ComfyUI server machine.

Trigger a render. Watch the ComfyUI console — you should see a workflow submitted and executed.

For a permanent setup, don’t enter these values per-clip — pre-fill them in each plugin’s defaults.json so every clip starts with the right server and paths. See Configuration & defaults.

Health check

A quick way to confirm everything is wired up correctly:

  1. In any web browser on the host machine, open http://<server-ip>:8188/. The ComfyUI UI should load.
  2. From a terminal on the host machine, write a test file to the share: echo hello > /<client-mount>/test.txt.
  3. From a terminal on the server, read it back: cat /<server-mount>/test.txt. Same file, same content.

If both work, the plugin will work.

Security notes

  • ComfyUI’s HTTP API has no authentication by default. Keep it on a trusted internal network and use firewall rules to restrict access.
  • The shared filesystem similarly has no plugin-level access control. Anyone with access to the share can read EXR data flowing through it.
  • Model weights are large. Download them from the official upstream sources listed on each plugin page, not from random mirrors.

Plugin code is BSD-3-Clause. Model weights are subject to their upstream licenses; see each plugin page. Images from upstream model authors are reproduced for documentation purposes under applicable fair-use / fair-dealing provisions with attribution; see the credits page. Built by MaGMa for Reepost Studio. Funded by CNC (Centre national du cinéma et de l'image animée).

This site uses Just the Docs, a documentation theme for Jekyll.